LingBot-VA — Galaxea A1 Plug Insertion EEF

This is a full-parameter fine-tune of robbyant/lingbot-va-base for a Galaxea A1 plug-insertion task using episode-relative end-effector (EEF) actions. The model jointly predicts video latents and robot actions.

The tokenizer, text encoder, and VAE are inherited unchanged from LingBot-VA base. transformer/ contains the fine-tuned step-500 weights.

Training data

The source dataset is pengyue-polaron/nyush-galaxea-a1-plug-insertion-eef-v21, a LeRobot Dataset v2.1 dataset for the instruction:

Pick up the charger and insert it into the first socket from the left on the power strip.

Field	Value
Available demonstrations	31 episodes / 24,638 frames at 30 FPS
Training split	26 episodes / 20,109 frames
Reserved split	Episodes `0, 7, 15, 23, 30` / 4,529 frames
Cameras	`front` 480×480 RGB; `wrist` 640×480 RGB
Action	Episode-relative EEF position + quaternion + normalized gripper

Training

Setting	Value
Optimizer steps	500
Hardware	2 × NVIDIA H100 80 GB
Distributed strategy	Full-parameter FSDP
Precision	bfloat16
Effective global batch size	16
Optimizer	Fused AdamW
Learning rate	1e-5 with 10-step warmup, then constant
Adam betas / weight decay	(0.9, 0.95) / 0.1
Temporal sampling	Random aligned clips, at most 64 latent frames
Objective	Video latent loss + action loss

For the last 100 optimizer steps, the mean video-latent loss was 0.115598 and the mean action loss was 0.004789.

Training run: Weights & Biases
Source repository: pengyue-polaron/lingbot-va-galaxea-a1
Starting source revision: 9fc8d8d0dfb5a4aee10d0d2fa184f911098492f9

The training working tree contained uncommitted experiment changes. The exact configs, dataset loader, trainer, and launch scripts used for this run are included in this model repository under configs/ and training/.

The eight source action values are mapped to LingBot-VA action channels [0, 1, 2, 3, 4, 5, 6, 28]; channel 28 carries the gripper command.

Evaluation status

The five reserved episodes were excluded from training, but this checkpoint has not yet been evaluated in closed-loop robot rollouts. No plug-insertion success rate is claimed. Training losses alone should not be interpreted as task success.

Repository contents

The repository contains the step-500 transformer, base tokenizer/text encoder/VAE, exact training files, and training_summary.json. Optimizer and other resumable training-state files are not included.